Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. A method for automated reconstruction of a 3D computed tomography (CT) volume one or more X-ray images, comprising: generating a sparse 3D volume from one or more X-ray images of a patient; and generating a final reconstructed 3D CT volume from the sparse 3D volume using a trained deep neural network.
This invention relates to automated 3D computed tomography (CT) volume reconstruction from X-ray images. The problem addressed is the time-consuming and resource-intensive nature of traditional CT reconstruction methods, which often require extensive computational power and manual intervention. The invention provides a faster, more efficient approach by leveraging deep learning techniques. The method involves two main steps. First, a sparse 3D volume is generated from one or more X-ray images of a patient. This sparse volume serves as an initial representation of the scanned object, capturing basic structural information with reduced data density. Second, a trained deep neural network processes this sparse volume to produce a final, high-quality 3D CT volume. The neural network is designed to refine and enhance the sparse data, filling in gaps and improving resolution to match the quality of traditionally reconstructed CT volumes. The deep neural network is trained on a dataset of known X-ray and CT image pairs, enabling it to learn the relationships between sparse and dense 3D representations. This approach reduces the need for extensive computational resources and manual adjustments, making the reconstruction process more efficient while maintaining accuracy. The method is particularly useful in medical imaging, where rapid and precise 3D reconstructions are critical for diagnosis and treatment planning.
2. The method of claim 1 , wherein the one or more X-ray images of the patient comprise a first x-ray image and a second x-ray image, and generating the sparse 3D volume from the one or more X-ray images of the patient comprises: generating the sparse 3D volume from the first X-ray image and the second X-ray image using a tomographic reconstruction algorithm.
This invention relates to medical imaging, specifically generating a sparse 3D volume from X-ray images of a patient. The problem addressed is the need for efficient and accurate 3D reconstruction from limited X-ray projections, which is useful in applications like intraoperative imaging or radiation therapy where real-time 3D visualization is required. The method involves acquiring at least two X-ray images of the patient from different angles. These images are then processed using a tomographic reconstruction algorithm to generate a sparse 3D volume. The sparse 3D volume is a simplified representation of the patient's anatomy, containing only essential structural information rather than a full, high-resolution 3D model. This approach reduces computational complexity and processing time while still providing sufficient detail for clinical applications. The tomographic reconstruction algorithm may include techniques such as filtered back projection or iterative reconstruction methods, which are optimized to work with minimal input data. The resulting sparse 3D volume can be used for various purposes, including guiding medical procedures, monitoring treatment progress, or planning further imaging or interventions. The method ensures that the 3D reconstruction is both rapid and accurate, making it suitable for time-sensitive medical scenarios.
3. The method of claim 2 , wherein the one or more x-ray images of the patient comprise only the first and second x-ray images, and generating the sparse 3D volume from the first X-ray image and the second X-ray image using a tomographic reconstruction algorithm comprises: generating the sparse 3D volume from the first X-ray image and the second X-ray image without any additional x-ray images using a tomographic reconstruction algorithm.
This invention relates to medical imaging, specifically generating a sparse 3D volume from a limited number of X-ray images. The problem addressed is the need for accurate 3D reconstruction of anatomical structures using fewer X-ray images than traditional methods, reducing radiation exposure and acquisition time. The method involves using only two X-ray images of a patient to generate a sparse 3D volume. The first and second X-ray images are processed using a tomographic reconstruction algorithm to produce the 3D volume without requiring any additional X-ray images. This approach minimizes radiation exposure while still providing sufficient 3D structural information for medical analysis. The tomographic reconstruction algorithm is designed to interpolate and reconstruct the 3D volume from the limited 2D projections, ensuring accuracy despite the reduced input data. The method is particularly useful in clinical settings where rapid imaging is needed, such as emergency rooms or intraoperative procedures, where minimizing patient exposure to radiation is critical. The sparse 3D volume can be used for diagnostic purposes, treatment planning, or guiding medical interventions. The algorithm compensates for the lack of additional images by leveraging advanced reconstruction techniques to infer missing data, ensuring the resulting 3D model is clinically useful.
4. The method of claim 1 , further comprising: generating a 3D segmentation mask of a target object from the sparse 3D volume using the trained deep neural network.
This invention relates to 3D object segmentation using deep neural networks. The problem addressed is the accurate identification and segmentation of target objects within sparse 3D volumes, which is challenging due to incomplete or noisy data. The solution involves training a deep neural network to process sparse 3D volumes and generate a 3D segmentation mask that delineates the target object. The trained network is applied to new sparse 3D data to produce the segmentation mask, enabling precise object localization and boundary definition. The method leverages deep learning to infer missing or ambiguous information in the sparse volume, improving segmentation accuracy compared to traditional approaches. The segmentation mask can be used for various applications, including object recognition, medical imaging, autonomous navigation, and industrial inspection. The invention enhances the reliability of 3D object detection in scenarios where data is incomplete or noisy, making it suitable for real-world applications where high precision is required.
5. The method of claim 4 , wherein the trained deep neural network is a multi-output deep image-to-image network having encoder layers that code the sparse 3D volume into a code whose size is smaller than the spare 3D volume and decoder layers that decode the code into the final reconstructed 3D volume and the 3D segmentation mask of the target object.
This invention relates to a deep learning-based system for reconstructing 3D volumes and generating segmentation masks from sparse 3D input data. The problem addressed is the efficient and accurate reconstruction of detailed 3D structures from limited or incomplete input data, which is common in medical imaging, computer vision, and other fields where high-resolution 3D data is costly or difficult to acquire. The system uses a multi-output deep image-to-image network, a type of neural network designed to process and transform 3D data. The network includes encoder layers that compress the sparse 3D input volume into a compact code representation, significantly reducing its size while preserving essential features. This compressed code is then processed by decoder layers, which reconstruct the final high-resolution 3D volume and simultaneously generate a 3D segmentation mask that identifies the target object within the volume. The segmentation mask is a binary or multi-class output that delineates the boundaries of the object, enabling precise localization and analysis. The approach leverages deep learning to improve reconstruction quality and segmentation accuracy compared to traditional methods, particularly when dealing with sparse or noisy input data. The multi-output design allows the network to perform both reconstruction and segmentation in a single, unified process, enhancing efficiency and reducing computational overhead. This method is particularly useful in applications such as medical imaging, where accurate 3D reconstructions and segmentations are critical for diagnosis and treatment planning.
6. The method of claim 1 , wherein the trained deep neural network is a deep image-to-image network that is trained in a generative adversarial network together with a discriminator network for distinguishing between synthesized reconstructed 3D CT volumes generated by the deep image-to-image network from input sparse 3D volume training samples and real reconstructed 3D CT volume training samples.
This invention relates to medical imaging, specifically improving the reconstruction of 3D computed tomography (CT) volumes from sparse input data. The core problem addressed is the challenge of generating high-quality 3D CT images from limited or incomplete input data, which is common in low-dose or fast-scanning CT applications. Traditional reconstruction methods often produce artifacts or lack detail due to insufficient data. The solution involves a deep image-to-image neural network trained within a generative adversarial network (GAN) framework. The GAN consists of two components: a generator (the deep image-to-image network) and a discriminator. The generator takes sparse 3D volume training samples as input and synthesizes reconstructed 3D CT volumes. The discriminator is trained to distinguish between these synthesized volumes and real reconstructed 3D CT volume training samples. Through adversarial training, the generator learns to produce high-fidelity 3D CT reconstructions that closely resemble real CT volumes, even from sparse input data. This approach enhances image quality while reducing artifacts, making it suitable for applications requiring rapid or low-dose CT imaging. The method leverages deep learning to improve reconstruction accuracy without requiring additional hardware or extensive data acquisition.
7. The method of claim 1 , wherein the trained deep neural network is a deep image-to-image network that is trained in a conditional-generative adversarial network together with a discriminator network for distinguishing between synthesized reconstructed 3D CT volumes generated by the deep image-to-image network from input sparse 3D volume training samples and real reconstructed 3D CT volume training samples, conditioned on the input sparse 3D volume training samples.
This invention relates to medical imaging, specifically improving the reconstruction of 3D computed tomography (CT) volumes from sparse input data. The problem addressed is the need for high-quality 3D CT reconstructions from limited or incomplete input data, which is common in medical imaging to reduce radiation exposure or scan time. Traditional reconstruction methods often produce artifacts or lack detail when input data is sparse. The solution involves a deep image-to-image neural network trained within a conditional generative adversarial network (GAN) framework. The network takes sparse 3D volume training samples as input and generates synthesized reconstructed 3D CT volumes. A discriminator network is used to distinguish between these synthesized volumes and real reconstructed 3D CT volume training samples, ensuring the generated outputs closely match real data. The training process conditions the generator on the input sparse volumes, improving reconstruction accuracy. This approach leverages adversarial training to enhance the realism and fidelity of the reconstructed CT images, making it particularly useful in medical diagnostics where high-quality imaging is critical. The method reduces the need for extensive input data while maintaining image quality.
8. The method of claim 7 , wherein the conditional-generative adversarial network is integrated with a voxel-wise cost function that computes a voxel-wise error between the synthesized reconstructed 3D CT volumes generated by the deep image-to-image network from input sparse 3D volume training samples and corresponding ground-truth reconstructed 3D CT volume training samples, and the deep image-to-image network and the discriminator network are trained together to optimize, over a plurality of training samples, a minimax objective function that includes a first term that calculates an error using the voxel-wise cost function, a second term that calculates an error of the discriminator network classifying the real reconstructed 3D CT training samples, and a third term that calculates and error of the discriminator network classifying the synthesized reconstructed 3D CT volumes generated by the deep image-to-image network.
This invention relates to medical imaging, specifically improving the reconstruction of 3D computed tomography (CT) volumes from sparse input data using deep learning techniques. The problem addressed is the challenge of generating high-quality 3D CT reconstructions from limited or incomplete input data, which is common in medical imaging due to factors like radiation dose reduction or hardware limitations. The solution involves a conditional-generative adversarial network (GAN) integrated with a voxel-wise cost function. The GAN consists of a deep image-to-image network that synthesizes reconstructed 3D CT volumes from sparse 3D volume training samples, and a discriminator network that distinguishes between real and synthesized CT volumes. The voxel-wise cost function computes the error between the synthesized volumes and corresponding ground-truth CT volumes at the voxel level, ensuring high-fidelity reconstructions. During training, the deep image-to-image network and discriminator network are jointly optimized using a minimax objective function. This function includes three terms: a voxel-wise error term, a discriminator error term for real CT samples, and a discriminator error term for synthesized CT samples. By balancing these terms, the network learns to generate realistic and accurate 3D CT reconstructions from sparse input data. This approach enhances image quality while reducing the need for extensive or high-dose imaging data.
9. An apparatus for automated reconstruction of a 3D computed tomography (CT) volume one or more X-ray images, comprising: means for generating a sparse 3D volume from one or more X-ray images of a patient; and means for generating a final reconstructed 3D CT volume from the sparse 3D volume using a trained deep neural network.
This invention relates to medical imaging, specifically the automated reconstruction of 3D computed tomography (CT) volumes from X-ray images. The problem addressed is the time-consuming and computationally intensive nature of traditional CT reconstruction methods, which often require extensive data processing and manual intervention. The invention provides a solution by leveraging deep learning to streamline the reconstruction process. The apparatus includes a system for generating a sparse 3D volume from one or more X-ray images of a patient. This sparse volume serves as an initial representation of the anatomical structure, capturing basic geometric and density information from the X-ray projections. The apparatus further includes a system for generating a final reconstructed 3D CT volume from the sparse 3D volume using a trained deep neural network. The neural network refines the sparse volume, enhancing details, correcting artifacts, and producing a high-quality 3D CT image. The deep learning approach reduces computational overhead and accelerates reconstruction compared to conventional methods. The invention is particularly useful in clinical settings where rapid and accurate 3D imaging is required for diagnosis and treatment planning.
10. The apparatus of claim 9 , further comprising: means for generating a 3D segmentation mask of a target object from the sparse 3D volume using the trained deep neural network.
This invention relates to 3D object segmentation in sparse volumetric data, addressing challenges in accurately identifying and isolating target objects within incomplete or low-resolution 3D volumes. The apparatus includes a trained deep neural network designed to process sparse 3D data, where sparsity refers to missing or incomplete volumetric information. The neural network is trained to infer missing data and generate a high-fidelity 3D segmentation mask of the target object. The segmentation mask delineates the object's boundaries within the sparse volume, enabling precise localization and shape reconstruction. The apparatus may also include preprocessing modules to enhance the sparse data quality before neural network processing, as well as post-processing steps to refine the segmentation mask. The system is particularly useful in applications like medical imaging, autonomous navigation, and industrial inspection, where sparse 3D data is common, and accurate object segmentation is critical. The deep neural network leverages learned features from training data to compensate for sparsity, improving segmentation accuracy compared to traditional methods that struggle with incomplete inputs. The invention focuses on automating the segmentation process while maintaining robustness across varying levels of data sparsity.
11. The apparatus of claim 9 , wherein the trained deep neural network is a deep image-to-image network that is trained in a generative adversarial network together with a discriminator network for distinguishing between synthesized reconstructed 3D CT volumes generated by the deep image-to-image network from input sparse 3D volume training samples and real reconstructed 3D CT volume training samples.
This invention relates to medical imaging, specifically to a system for reconstructing high-quality 3D computed tomography (CT) volumes from sparse input data using deep learning. The problem addressed is the computational and time-intensive nature of traditional CT reconstruction methods, which often require extensive data acquisition to produce accurate 3D images. The solution involves a deep image-to-image neural network trained within a generative adversarial network (GAN) framework. The GAN consists of two components: a generator network that synthesizes reconstructed 3D CT volumes from sparse 3D input samples, and a discriminator network that distinguishes between the synthesized volumes and real reconstructed 3D CT volumes. During training, the generator learns to produce realistic 3D CT reconstructions by minimizing the discriminator's ability to differentiate between real and synthetic data. This approach enables faster and more efficient CT reconstruction while maintaining high image quality, reducing the need for extensive data acquisition. The system is particularly useful in medical imaging applications where rapid and accurate 3D reconstructions are critical.
12. The apparatus of claim 9 , wherein the trained deep neural network is a deep image-to-image network that is trained in a conditional-generative adversarial network together with a discriminator network for distinguishing between synthesized reconstructed 3D CT volumes generated by the deep image-to-image network from input sparse 3D volume training samples and real reconstructed 3D CT volume training samples, conditioned on the input sparse 3D volume training samples.
This invention relates to medical imaging, specifically improving the reconstruction of 3D computed tomography (CT) volumes from sparse input data. The problem addressed is the need for high-quality 3D CT reconstructions when only limited or incomplete input data is available, which is common in low-dose or fast-scanning scenarios. Traditional reconstruction methods often produce artifacts or require extensive computational resources. The solution involves a deep image-to-image network trained within a conditional generative adversarial network (GAN) framework. The deep neural network takes sparse 3D volume data as input and generates a synthesized, high-quality 3D CT volume. A discriminator network is trained simultaneously to distinguish between the synthesized volumes and real, high-quality 3D CT volumes, ensuring the generated output closely matches real data. The training process is conditioned on the input sparse volumes, meaning the network learns to reconstruct missing data while preserving anatomical structures. This approach leverages adversarial training to enhance reconstruction quality, reducing artifacts and improving diagnostic accuracy. The method is particularly useful in medical imaging where reducing radiation exposure or scan time is critical.
13. The apparatus of claim 12 , wherein the conditional-generative adversarial network is integrated with a voxel-wise cost function that computes a voxel-wise error between the synthesized reconstructed 3D CT volumes generated by the deep image-to-image network from input sparse 3D volume training samples and corresponding ground-truth reconstructed 3D CT volume training samples, and the deep image-to-image network and the discriminator network are trained together to optimize, over a plurality of training samples, a minimax objective function that includes a first term that calculates an error using the voxel-wise cost function, a second term that calculates an error of the discriminator network classifying the real reconstructed 3D CT training samples, and a third term that calculates and error of the discriminator network classifying the synthesized reconstructed 3D CT volumes generated by the deep image-to-image network.
This invention relates to medical imaging, specifically improving the reconstruction of 3D computed tomography (CT) volumes from sparse input data using deep learning techniques. The problem addressed is the challenge of generating high-quality 3D CT reconstructions from limited or incomplete input data, which is common in low-dose or fast-scanning scenarios where radiation exposure or acquisition time must be minimized. The apparatus integrates a conditional-generative adversarial network (cGAN) with a voxel-wise cost function to enhance the accuracy of 3D CT volume reconstruction. The cGAN consists of a deep image-to-image network and a discriminator network. The image-to-image network generates synthesized 3D CT volumes from sparse input training samples, while the discriminator network evaluates the realism of these synthesized volumes against ground-truth CT volumes. The voxel-wise cost function computes the error between the synthesized and ground-truth volumes at each voxel, ensuring fine-grained accuracy. During training, the deep image-to-image network and discriminator network are jointly optimized using a minimax objective function. This function includes three terms: the first term measures the voxel-wise reconstruction error, the second term assesses the discriminator's ability to classify real CT volumes, and the third term evaluates the discriminator's performance in classifying synthesized volumes. By balancing these terms, the network learns to produce high-fidelity 3D CT reconstructions from sparse input data, improving diagnostic accuracy while reducing radiation exposure or scan time.
14. A non-transitory computer-readable medium storing computer program instructions for automated reconstruction of a 3D computed tomography (CT) volume one or more X-ray images, the computer program instructions when executed by a processor cause the processor to perform operations comprising: generating a sparse 3D volume from one or more X-ray images of a patient; and generating a final reconstructed 3D CT volume from the sparse 3D volume using a trained deep neural network.
This invention relates to medical imaging, specifically automated reconstruction of 3D computed tomography (CT) volumes from X-ray images. Traditional CT reconstruction methods require multiple X-ray projections and are computationally intensive. The invention addresses this by enabling high-quality 3D CT volume reconstruction from fewer X-ray images using deep learning techniques. The system generates a sparse 3D volume from one or more X-ray images of a patient. This initial sparse volume is a preliminary representation with limited detail. The system then processes this sparse volume through a trained deep neural network to produce a final, high-resolution 3D CT volume. The neural network is designed to enhance and refine the sparse volume, filling in missing data and improving image quality. The approach reduces the number of X-ray images needed while maintaining diagnostic accuracy, lowering radiation exposure for patients and computational demands. The deep neural network is pre-trained on a dataset of X-ray images and corresponding CT volumes, allowing it to learn the mapping between sparse and high-quality 3D reconstructions. This method is particularly useful in clinical settings where rapid, high-quality imaging is required with minimal radiation exposure. The invention improves upon traditional iterative reconstruction techniques by leveraging deep learning for faster and more accurate 3D volume generation.
15. The non-transitory computer-readable medium of claim 14 , wherein the one or more X-ray images of the patient comprise a first x-ray image and a second x-ray image, and generating the sparse 3D volume from the one or more X-ray images of the patient comprises: generating the sparse 3D volume from the first X-ray image and the second X-ray image using a tomographic reconstruction algorithm.
This invention relates to medical imaging, specifically the generation of a sparse 3D volume from X-ray images of a patient. The problem addressed is the need for efficient and accurate 3D reconstruction from limited X-ray data, which is often required in clinical settings where full tomographic scans are impractical or unnecessary. The invention involves a non-transitory computer-readable medium storing instructions for processing X-ray images to create a sparse 3D volume. The system uses at least two X-ray images of the patient, referred to as a first X-ray image and a second X-ray image. These images are processed using a tomographic reconstruction algorithm to generate the sparse 3D volume. The algorithm reconstructs a three-dimensional representation of the patient's anatomy from the two-dimensional X-ray projections, producing a volume that is less dense than a full tomographic scan but still provides useful structural information. The sparse 3D volume is derived from the first and second X-ray images, which may be acquired from different angles or positions to enhance reconstruction accuracy. The tomographic reconstruction algorithm may include techniques such as filtered back projection, iterative reconstruction, or machine learning-based methods to optimize the 3D model from the limited input data. This approach reduces radiation exposure and computational requirements compared to traditional CT scans while still enabling 3D visualization for diagnostic or procedural guidance.
16. The non-transitory computer-readable medium of claim 14 , wherein the operations further comprise: generating a 3D segmentation mask of a target object from the sparse 3D volume using the trained deep neural network.
This invention relates to computer vision and 3D object segmentation, specifically addressing the challenge of accurately segmenting target objects in sparse 3D volumes using deep neural networks. The technology involves a trained deep neural network that processes sparse 3D volume data to generate a 3D segmentation mask of a target object. The sparse 3D volume may be derived from various sources, such as LiDAR scans, medical imaging, or other depth-sensing technologies, where data points are not densely packed. The deep neural network is trained to interpret these sparse data points and produce a precise 3D mask that delineates the boundaries of the target object within the volume. This segmentation mask can be used for applications like object detection, medical imaging analysis, autonomous navigation, or industrial inspection, where accurate 3D object delineation is critical. The method leverages the network's ability to infer missing data from sparse inputs, improving segmentation accuracy in scenarios where traditional dense 3D data is unavailable or impractical to acquire. The invention enhances the efficiency and reliability of 3D object segmentation in real-world applications where data sparsity is a common constraint.
17. The non-transitory computer-readable medium of claim 16 , wherein the trained deep neural network is a multi-output deep image-to-image network having encoder layers that code the sparse 3D volume into a code whose size is smaller than the spare 3D volume and decoder layers that decode the code into the final reconstructed 3D volume and the 3D segmentation mask of the target object.
This invention relates to a computer-implemented method for reconstructing 3D volumes and generating segmentation masks using a deep neural network. The technology addresses the challenge of efficiently processing sparse 3D data, which is common in medical imaging, autonomous navigation, and other fields where high-resolution 3D reconstructions are needed but computational resources are limited. The system employs a multi-output deep image-to-image network designed to process sparse 3D volumes. The network includes encoder layers that compress the input sparse 3D volume into a compact code representation, significantly reducing its size while preserving essential features. This encoded representation is then passed through decoder layers, which reconstruct the final 3D volume and simultaneously generate a 3D segmentation mask of the target object. The segmentation mask identifies and delineates the object within the reconstructed volume, enabling precise analysis or further processing. The network is trained to optimize both reconstruction accuracy and segmentation performance, ensuring that the output is both structurally accurate and semantically meaningful. This approach improves efficiency by avoiding separate reconstruction and segmentation steps, reducing computational overhead while maintaining high-quality results. The method is particularly useful in applications requiring real-time or near-real-time 3D data processing, such as medical diagnostics, robotic vision, and autonomous systems.
18. The non-transitory computer-readable medium of claim 14 , wherein the trained deep neural network is a deep image-to-image network that is trained in a generative adversarial network together with a discriminator network for distinguishing between synthesized reconstructed 3D CT volumes generated by the deep image-to-image network from input sparse 3D volume training samples and real reconstructed 3D CT volume training samples.
This invention relates to medical imaging, specifically improving the reconstruction of 3D computed tomography (CT) volumes from sparse input data using deep learning techniques. The problem addressed is the need for high-quality 3D CT reconstructions from limited or incomplete input data, which is common in medical imaging to reduce radiation exposure or scan time. Traditional reconstruction methods often produce artifacts or require extensive computational resources. The solution involves a trained deep neural network, specifically a deep image-to-image network, designed to generate high-fidelity 3D CT volumes from sparse input data. This network is trained using a generative adversarial network (GAN) framework, which includes a discriminator network. The discriminator network is trained to distinguish between synthesized 3D CT volumes generated by the deep image-to-image network and real 3D CT volume training samples. The adversarial training process ensures that the generated 3D CT volumes closely resemble real CT volumes, improving reconstruction quality. The deep image-to-image network processes sparse 3D volume training samples to produce reconstructed 3D CT volumes, while the discriminator network evaluates the realism of these reconstructions. This approach enhances the accuracy and fidelity of 3D CT reconstructions, making it particularly useful in medical imaging applications where high-quality images are critical for diagnosis and treatment planning.
19. The non-transitory computer-readable medium of claim 14 , wherein the trained deep neural network is a deep image-to-image network that is trained in a conditional-generative adversarial network together with a discriminator network for distinguishing between synthesized reconstructed 3D CT volumes generated by the deep image-to-image network from input sparse 3D volume training samples and real reconstructed 3D CT volume training samples, conditioned on the input sparse 3D volume training samples.
This invention relates to medical imaging, specifically improving the reconstruction of 3D computed tomography (CT) volumes from sparse input data. The problem addressed is the need for high-quality 3D CT reconstructions when only limited or incomplete input data is available, which is common in low-dose or fast-scanning scenarios. Traditional reconstruction methods often produce artifacts or require extensive computational resources. The solution involves a deep image-to-image network trained within a conditional generative adversarial network (GAN) framework. The deep neural network generates synthesized 3D CT volumes from sparse 3D input samples, while a discriminator network distinguishes between these synthesized volumes and real reconstructed 3D CT volumes. The training process conditions the generator on the sparse input data, ensuring the output closely matches real CT volumes. This approach leverages adversarial training to enhance reconstruction quality, reducing artifacts and improving fidelity. The method is particularly useful in medical imaging where high-quality reconstructions are critical for diagnosis and treatment planning, while minimizing radiation exposure or scan time. The trained network can be deployed to process new sparse 3D volume inputs, producing accurate and detailed 3D CT reconstructions.
20. The non-transitory computer-readable medium of claim 19 , wherein the conditional-generative adversarial network is integrated with a voxel-wise cost function that computes a voxel-wise error between the synthesized reconstructed 3D CT volumes generated by the deep image-to-image network from input sparse 3D volume training samples and corresponding ground-truth reconstructed 3D CT volume training samples, and the deep image-to-image network and the discriminator network are trained together to optimize, over a plurality of training samples, a minimax objective function that includes a first term that calculates an error using the voxel-wise cost function, a second term that calculates an error of the discriminator network classifying the real reconstructed 3D CT training samples, and a third term that calculates and error of the discriminator network classifying the synthesized reconstructed 3D CT volumes generated by the deep image-to-image network.
This invention relates to medical imaging, specifically improving the reconstruction of 3D computed tomography (CT) volumes from sparse input data. The problem addressed is the challenge of generating high-quality 3D CT reconstructions from limited or incomplete input data, which is common in low-dose or fast-scanning scenarios. Traditional methods often produce artifacts or lose fine details. The solution involves a deep learning-based approach using a conditional-generative adversarial network (GAN) integrated with a voxel-wise cost function. The system includes a deep image-to-image network that synthesizes reconstructed 3D CT volumes from sparse 3D volume training samples. A discriminator network evaluates the realism of these synthesized volumes by comparing them to ground-truth reconstructed 3D CT volumes. The voxel-wise cost function computes the error between the synthesized and ground-truth volumes at each voxel, ensuring precise reconstruction. The deep image-to-image network and discriminator network are jointly trained to optimize a minimax objective function. This function combines three terms: a voxel-wise error term, a discriminator error term for real samples, and a discriminator error term for synthesized samples. The training process iteratively refines the networks to minimize reconstruction errors while improving the discriminator's ability to distinguish real from synthesized volumes. This approach enhances the quality and accuracy of 3D CT reconstructions from sparse input data.
Unknown
July 14, 2020
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.